Empirical Indicators of Crisis Phase in the Middle East, 1979-1995

A number of studies of crisis behavior—for example, the Butterworth, SHERFACS, and CASCON data sets—have assumed that political behavior goes through a series of clear "phases" characterized by distinct patterns of interactions. To date, these phases have been identified contextually by human coders rather than through any systematic procedures. This paper uses event data to analyze phases in the political behavior between eight Middle Eastern actors—Egypt, Israel, Jordan, Lebanon, the Palestinians, Syria, the United States and USSR/Russia—for the period July 1979 to June 1995. The event data are generated from the Reuters news service using the KEDS machine-coding system, then converted to monthly timeseries using the Goldstein (1992) scaling. The system is then analyzed in SPSS using three datareduction methods: • factor analysis of the multiple time series to identify the underlying patterns of behavior found in the system; • discriminant analysis to determine whether the dyadic time-series can predict the phases assigned by a human coder; • cluster analysis of behavior vectors over time to identify inductively the phases. As we expected, the factor analysis identifies the Israeli-Palestinian conflict and the war in Lebanon as the two most important features of the data; other factors reflect inter-Arab and major-power relationships. The factor scores do not, however, prove very useful in either the discriminant or cluster analysis. The discriminant analysis identified the human-coded phases with about 90% accuracy; a stepwise discriminant did this with 70% accuracy using only 12 of the 54 dyads. Finally, the cluster analysis identified five distinct phases that align fairly well with the human-coded phases, particularly in the first half of the time period. The overall results of the analysis correspond fairly well with our contextual understanding of the political situation in the Middle East and show considerable promise in providing a means of studying the behavior of a complex N-actor system systematically, rather than focusing on a small number of dyads. The paper concludes with observations on how these methods might be applied to the problems of early warning. Introduction A number of contemporary studies of crises assume that political behaviors go through a series of phases that are delineated by an emphasis on different sets of behavior. In the empirical literature, crisis "phase" has been explicitly coded in data sets such as the Butterworth international dispute resolution dataset (Butterworth 1976), CASCON (Bloomfield and Moulton 1989) and SHERFACS (Sherman and Neack 1993).1 Describing the early CASCON work, Sherman and Neack explain that: ...conflict is seen "as a sequence of phases". Movement from phase to phase in a conflict occurs as "the factors interact in such a way as to push the conflict ultimately across a series of thresholds toward or away from violence (Bloomfield and Leiss 1969). Characteristics of disputes can be visualized as the timing and sequencing of movement between and among phases. Processes of escalation of violence, resolution or amelioration of the seriousness (threat of violence-hostilities) and settlement are identifiable through the use of phrase structures. (Sherman and Neack 1993:90) SHERFACS, for example, codes six phases: "dispute phase," "conflict phase," "hostilities phase," "post-hostilities conflict phase," post-hostilities dispute phase," and "settlement phase". In the policy literature, crisis phase has emerged as a key part of the "preventive diplomacy" concept, because of the assumption that diplomacy can be more effective in the early stages of a crisis (e.g., before the outbreak of military hostility) than in later periods. In situations where preventive diplomacy is not an option, crisis phase may still be of utility in providing an early warning of, for example, large-scale refugee movements. Depending on the phase of a crisis, a localized outbreak of military action may be contained without generating large numbers of refugees, or it might rapidly spread, requiring the need for an international response. Much of the literature on ethnic conflict assumes that militarized ethnic disputes such as those found in the former Yugoslavia, the former Soviet Union, Rwanda, Sri Lanka and other locations go through a series of relatively predictable phases (see Alker, Gurr, and Rupesinghe 1995). The phases identified in the Butterworth, CASCON and SHERFACS datasets have all been assigned by human coders. While human-coding is obviously necessary in the early stages of the development of a new concept, it presents three problems. First, when the coding of a phase is 1 Sherman and Neack (1993) provide a review of the evolution of these data sets. dependent on human judgment, the de facto definition of the phase is likely to drift over time. This can occur as a single coder becomes more experienced with the data, and is very likely to occur during attempts to transfer the definitions across projects. Consequently, a phase variable measured across data sets may have significant measurement error, and phases in one data set may seem to have different implications than phases in another because the coders were, in fact, working with disparate definitions. In contrast, the statistical identification of phases—combined with the machine-coding of event data (Gerner et al. 1994)—should make it possible to code crisis phase consistently and efficiently in a variety of contexts and from an assortment of different sources. Second, in the earlier research, phases were coded retrospectively. While this may be useful from the standpoint of validity, it is likely to exaggerate the effectiveness of phase as an early warning indicator, as well as overestimating the ability of a policy analyst to assess the current phase of a crisis while experiencing it. If phase structure is going to be developed as an effective tool, it must be able to work in real time rather than with "20-20 hindsight." Third, the tendency of human analysts to impose order on episodes of political events means that in some instances human coders may see phases that are not actually present in the data. This is a problem for the development of statistical early warning indicators because either (a) the human coder is correctly identifying the phase, but is making the assessment based on additional information that must be provided to the statistical early warning system (i.e., any model using phase variables coded differently will have a specification error) or (b) the human coder is incorrectly identifying the phase, which will bias any statistical estimates made with the data. We suspect that human-coded phase identification contains both types of error. This paper analyses the phase structure of political events in the Levant—Egypt, Israel, Jordan, Lebanon, the Palestinians, Syria—for the period 1979-1995 using event data.2 This region exhibits some of the most complicated political behavior in the world, with a variety of state and non-state actors vying for influence in the context of the ongoing Arab-Israeli conflict and, until 1990, USA-Soviet competition. Unlike most of the human-coded work on crisis phase, this region—and the event data describing it—involves not a single crisis but several inter-linked disputes. The two dominant political themes have been the Israeli-Palestinian conflict and the Lebanese civil war—both going through phases of hostility and mediation 2 The advantages and disadvantages of event data are discussed by Azar, Brody and McClelland (1972), Burgess and Lawton (1972), Azar and Ben Dak (1975), Daly and Andriole (1980), Doran, Pendley, and Antunes (1973), Azar and Ben-Dak (1975), Peterson (1975), Munton (1978), Goldstein and Freeman (1990), Merritt, Muncaster, and Zinnes (1993), Gerner et al. (1994), and Schrodt (1995). Marlin-Bennet and Roberts (1993) provide a recent discussion of event data from a research perspective. during this period—but there were also other key focal points, such as US efforts on the resolution of the larger Arab-Israeli dispute and spin-off interactions from the Iran-Iraq War and the Iraq's invasion of Kuwait. This extensive foreign policy activity presents a realistic challenge to any effort to identify crisis phase through statistical indicators because of the quantity and variety of material. Methodology One of the most substantial challenges in using event data to analyze behavior in an international subsystem is the sheer quantity of information available. An eight-actor system such as the one we are analyzing—the six Levantine actors plus the two superpowers—has event measures for 56 directed dyads, each contributing something to the overall behavior of the system as a whole. However, much of this behavior is inter-correlated because it is generated by a small set of political issues. In order to characterize the behavior of the system, we therefore need to do two things: 1. Ascertain the underlying issues that are generating the observed behaviors; 2. Identify the contributions of various dyads to determining those behaviors. If the concept of crisis phase is correct, we would expect that the observed behaviors would fall into distinct patterns over time and that we could determine the event behaviors that are characteristic of any phase. Finally, if the behaviors determining a phase are distinct and reflected in event data, it should also be possible to determine those phases inductively by looking at the data itself, without prior knowledge of the phases. The approach we will use to analyze the system is to look at the behavior of the vector [AB, AC, AD, ... , AH, BA, BC, ..., BH, CA, ...., HF, HG]t where A,B,...,H are the actors in the system and XYt is the total Goldstein-scaled events directed from X to Y aggregated over a month.3 Conceptually, the behavior of the system is simply the path that this vector traces over time in a 54-dimensional space.4 The high dimensionality of that space makes this path somewhat difficult to visualize. This is not a new problem in event data analysis: The response of most of the earlier event data 3 In other words, we converted each X∅Y event to its numerical score on the Goldstein scale, then totaled these numerical scores by month. Schrodt and Gerner (1994) gives a number of time series plots of the data for the 1982-1993 period. 4 We have excluded the USA->USR and USR->USA dyads from our analysis since most of their interactions did not deal with the Middle East. studies was to ignore N-actor systems and instead focus either on a small number of dyads (e.g., Ward 1982, Dixon 1986, Goldstein and Freeman 1990), or look at the interactions of one actor with a number of states (e.g., Howell and Barnes 1993). In situations where there are a small number of clearly dominant dyads—for example the USA/USSR/PRC triad of the Cold War—this is effective, but it is clearly not sufficient in a complex system such as the Middle East. The first objective of our analysis, therefore, is to see whether we can substantially decrease the dimensions of the behavior. Reducing the dimensionality will be possible provided there are consistent correlations between some of the dyadic behaviors in the system. There are at least three substantive reasons that we can expect this to be the case. First, there is considerable policy coordination between some of the states. In the extreme case, Lebanon's foreign policy has largely been run by Syria since the late 1980s. Lebanon's reported behaviors can be expected largely to mirror those of Syria, so knowledge of Syria's position may by itself provide sufficient information to predict Lebanon's position. The less extreme example of this would be simple policy coordination: For instance, during the Reagan years US and Israeli policies closely paralleled each other on most issues, as did those of Syria and the USSR. Second, all of the states in the system are reacting to the same set of events: Israel's invasion of Lebanon, Syria's eventual establishment of military hegemony in Lebanon, the Palestinian intifada, the Madrid peace process and so forth. To the extent that states share similar policy positions, they will react to these events in comparable ways. Furthermore, to the extent that certain issues receive greater emphasis in the Reuters reports—the source of our event data—we should expect to see correlated behavior related to those events. Finally, it is likely that some of the actors in the system have very little influence on the overall events. In our data set, two likely candidates are Egypt—which was diplomatically isolated during most of the period due to the Camp David agreements—and Jordan, which is comparatively small and neither initiates nor receives many events. A model of the behavior of the system could therefore ignore these (or other) actors. While there are a variety of different methods that could be used to reduce dimensionality through correlation, we will focus on the oldest and most well-understood: factor analysis (Kim and Mueller 1978). Factor analysis—which clusters variables based on their mutual correlations—allows us to answer both of the questions presented above. First, if foreign policies are primarily being determined by some exogenous events—whether the foreign policy of another actor or the interactions of another dyads (e.g., the Israeli-Palestinian conflict)—then these should show up as a distinct factor; further, the nature of that factor should be apparent from the dyads with which it is most strongly correlated. Second, if a state is not actively involved in the system (or is uninvolved with some subset of issues), this will be evident from an absence of correlation between its behavior and those factors. Finally, if a small number of factors explain a large amount of the variance in the system, models dealing with the systemic behavior can use those factors rather than the larger number of dyadic behaviors; the factors have the additional advantage of being uncorrelated (orthogonal). Conceptualizing the system as moving in a high-dimensional vector space also allows us to deal formally with the issue of crisis phase. In the vector terminology, a "phase" is characterized by a region in the vector space where points cluster over time. Empirically, a phase typology would be evident by the system spending most of its time inside the distinct clusters of behaviors that characterize the phase, with brief transitions between these clusters. Figure 1 illustrates this process informally for the World War II period, using the two dimensions of "talking versus fighting" and "local versus global involvement." The years prior to 1936 involved little violent conflict. The system then shifted to a series of militarized crises during the period 1936-38, and erupted into a full-scale European war in 1939-40. After a lull in the early part of 1941, the war spread first to the USSR, and then to the Pacific, and the 19421944 was characterized by a global war. In 1945, this war ended, first in Europe and then in the Pacific, but the post-war politics, rather than returning to the unilateralism/isolationism of the pre-war period, remained global. The 1946-47 cluster continues to characterize the system for most of the Cold War, with occasional departures from this cluster to take in Korea, Suez, the Cuban Missile Crisis and so forth. Figure 1 Schematic Representation of Phases during the WWII Period